Identification of prognostic markers related to homologous recombination deficiency in cholangiocarcinoma using CoxBoost and LASSO machine learning techniques

Yan Liu¹Cheng Zhou¹Tianhao Shen¹Xue Yu¹Qiuying Li¹Tinghui Jiang¹Wei Li^2*Yongqiang Zhu^1*

• ¹Oncology Intervention Department, Putuo Hospital, Shanghai University of Traditional Chinese Medicine, Shanghai, China
• ²Hepatopancreatobiliary Surgery Department, Putuo Hospital, Shanghai University of Traditional Chinese Medicine, Shanghai, China

Background: Cholangiocarcinoma (CHOL) is a highly aggressive malignancy with a poor prognosis. Homologous recombination deficiency (HRD) is associated with genomic instability and cancer progression, making it a potential therapeutic target. The aim ofthis study is to develop novel potential prognostic biomarkers and construct an HRD-based prognostic risk predictionmodel for CHOL to enhance clinical precision medicine. Methods: We analyzed HRD across cancers using multiple datasets including TCGA-CHOL, TCGA-LIHC, GDC TARGET-OS, and IMvigor210. HRD scores were calculated using data from Thorsson et al and the maftools R package was used for mutation data visualization and tumor mutational burden (TMB) calculation. Differential gene expression analysis identified HRD-related genes, validated in tumor and adjacent non-tumor tissues using RT-PCR. 10 machine-learning algorithms including RSF, LASSO, GBM, Survival-SVM, SuperPC, Ridge, plsRcox, CoxBoost, Stepwise Cox, Enet were selected to construct a prognostic model and validated in the E-MTAB-6389 and GSE107943. Among them, RSF, LASSO, CoxBoost and Stepwise Cox have the functions of dimension reduction and variable screening. Results: Comparative analysis demonstrated significant associations between HRD scores and genomic instability markers. High HRD scores independently predicted poorer overall survival (log-rank p = 0.043) and progression-free interval (log-rank p = 0.028). Immune infiltration analysis revealed higher levels of active B cells and regulatory T cells in the low-risk group, suggesting differential immune landscapes between risk groups. We identified the CoxBoost and LASSO algorithms as the optimal combination for creating a CoxBoost+ LASSO prognostic model. Using this model, we identified six genes (ANXA2P1, BBOX1, KLHL33, MN1, OR51A4, and TRDN) with significant differential expression. Conclusions: Our HRD-based prediction model offers a reliable tool for CHOL prognosis, suggesting new potential for six candidate genes as prognostic biomarkers. It highlights potential therapeutic targets and drug sensitivities, providing new insights into personalized treatment strategies for CHOL management.